Natural Language Processing

The ability to automatically, consistently, and correctly understand (and extract) information from textual sources is a key characteristic of many real-world AI applications. This trait is similarly critical in the Human Resources (HR) domain. Current state-of-the-art large pre-trained language models have recently demonstrated impressive performance on a wide range of NLP tasks, including natural-language generation, summarization, question answering, reading comprehension, and named entity recognition/resolution. But they have also shown limitations in areas like interpretability, controllability, transparency, and fairness.

At Megagon Labs we focus on how to take advantage of large pre-trained language models and go beyond the current state of the art. We work on the investigation, proposal, and deployment of new models, systems, and approaches that boost natural language processing capabilities. We do this by defining new architectures, using hybrid neuro-symbolic paradigms, and exploring domain-specific characteristics that positively impact the quality, consistency, fairness, and truthfulness of our solutions on HR and related domains.

Recent Publications:

Reasoning Capacity in Multi-Agent Systems: Limitations, Challenges and Human-Centered Solutions

Retrieval Helps or Hurts? A Deeper Dive into the Efficacy of Retrieval Augmentation to Language Models

Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions

Characterizing Large Language Models as Rationalizers of Knowledge-intensive Tasks

XATU: A Fine-grained Instruction-based Benchmark for Explainable Text Updates

Less is More for Long Document Summary Evaluation by LLMs

Human-LLM Collaborative Annotation Through Effective Verification of LLM Labels

MEGAnno+: A Human-LLM Collaborative Annotation System

Related Projects:

Coop: Convex Aggregation for Opinion Summarization

We developed Coop, a tool that enables us to generate more specific summaries by finding better summary vector in the latent space.

CoCoSum: Contrastive Summary for Two Comparable Entities

We developed a novel decoding algorithm, co-decoding. For the distinctive opinion summary generation, it emphasizes the distinctive words by contrasting the token probability distribution of the target entity against that of the counterpart entity. For the common opinion summary generation, it highlights the entity-pair specific words by aggregating token probability distributions.


GiNZA is an open-source Japanese NLP library with features such as a one-step installer, high-speed and high-precision analysis, and international capabilities for sentence structure analysis.